Large Model Applications in Automotive Industrialization: Practices, Architecture, and Case Studies

Overview – The session focuses on the practice and application of large models in automotive industrialization, covering five parts: ChatGPT development, large‑model fundamentals, empowerment of new industrialization, practical exploration, and a Q&A segment.

1. ChatGPT Development – Since OpenAI’s 2018 GPT release, models evolved from GPT‑1 to GPT‑3 with limited performance, while GPT‑3.5 achieved strong NLP capabilities. GPT‑4 introduced multimodal input, enabling image‑based tasks, role‑play, chart analysis, coding, and professional exam solving.

2. Large‑Model Fundamentals – Differences between BERT (encoder‑only) and GPT (decoder‑only) are highlighted, including network structure, usage (CLS embedding vs. conversational interaction), and pre‑training objectives (masked language modeling vs. next‑token prediction). InstructGPT improvements (human‑preferred dialogue data and reinforcement learning) further align ChatGPT responses with human preferences.

3. Empowering New Industrialization – Large models enable a shift from AI‑1.0 (small models) to AI‑2.0 (large models) by reducing data labeling effort and supporting cross‑domain tasks such as visual inspection, acoustic anomaly detection, and predictive analytics. Three application paradigms are described: instruction prompting, decision‑making assistance, and autonomous decision making.

4. Practice and Exploration in Industry – NIO’s AI platform is built on a layered architecture (chip, framework, model, service, business). Data and model are tightly coupled, allowing natural‑language data access. The platform implements three closed‑loops: data loop (ETL, annotation, storage, training, feedback), model loop (corpus, training, evaluation, A/B testing, deployment), and agent loop (continuous enhancement of AI agents). Intelligent quality inspection combines vision, audio, and digital analysis, while a cloud‑edge hybrid ensures real‑time detection on the shop floor.

5. Case Studies – (a) Three‑dimensional intelligent inspection (visual, acoustic, digital) reduces reliance on large labeled datasets. (b) G8D Agents map the eight‑step problem‑solving process to eight specialized agents, enabling rapid, automated quality issue resolution and knowledge capture. (c) Cloud‑edge architecture supports continuous model retraining and deployment.

Q&A Highlights – Discussed construction of G8D agents, knowledge storage (vector DB and Elasticsearch), and intent‑recognition optimization through multi‑level closed‑loop iteration.

Overall, the talk demonstrates how large language models and AI agents can transform automotive manufacturing through scalable, data‑driven, and autonomous solutions.